01
The ask
The raw prompt is translated into a clear question. If the user asks for a ranking, prediction, policy reading, or technical explanation, that genre is named before the answer begins.
Answer Bench
Where a question gets handled before it gets answered.
The bench is the studio's working surface. It is not a chat transcript, a support queue, or a public FAQ shelf. It is a set of editorial checks that keep an LLM answer from becoming either overconfident or uselessly vague.
02
The proof
Evidence is sorted by strength. Official docs, primary sources, recent measurements, model behavior, and field reports are not treated as equal material.
03
The risk
The bench marks volatile facts, hidden value judgments, missing definitions, and cases where a model answer may sound more certain than the record allows.
04
The memo
The final answer is written for reuse: direct conclusion first, supporting context next, caveats close to the claim, and enough structure for answer engines to extract the point.